基于深度学习语义分割模型的高分辨率遥感图像水体提取  

Water information extraction from high-resolution remote sensing images using the deep-learning based semantic segmentation model

作　　者：沈骏翱 马梦婷 宋致远 柳汀洲 张微 SHEN Jun’ao;MA Mengting;SONG Zhiyuan;LIU Tingzhou;ZHANG Wei(School of Software Technology,Zhejiang University,Ningbo 315048,China;School of Computer Science and Technology,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]浙江大学软件学院,宁波315048 [2]浙江大学计算机科学与技术学院,杭州310027

出　　处：《自然资源遥感》2022年第4期129-135,共7页Remote Sensing for Natural Resources

基　　金：浙江省重点研发计划项目“基于大数据的时空信息平台系统建设”(编号:2021C01031);宁波市自然科学基金项目“基于时空大数据和AIoT技术的污泥专运溯源管理系统研发与应用”(编号:2022S125)共同资助。

摘　　要：水体提取是高空间分辨率遥感影像应用中重要研究方向之一。传统识别方法仅利用水体的浅层特征,为了更好地挖掘遥感影像的深度信息,从而提升水体提取算法的鲁棒性,提高分割精度,提出了一种基于深度学习语义分割模型的水体提取方法。利用深度神经网络挖掘高分辨率遥感影像信息,同时引入注意力模块,整合深层信息与浅层地物的形状、结构、纹理和色调等信息,拟建立比现有模型具有更高准确率、更快预测速度的全新深度语义分割模型。最后,和传统识别方法以及常见语义分割模型进行对比消融实验。实验证明所提出算法模型的总体精度和效率均优于现有方法,且算法参数设置简单,受人工干预少。文章证明了深度学习以及注意力机制在高分辨率遥感影像水体提取任务上的准确性和高效性,提供了一种使用深度学习方法解决高分辨率遥感影像分割任务的可能,并对未来进行了展望。Water information extraction is an important study direction in the application of high spatial resolution remote sensing images.Conventional recognition methods only focus on the shallow features of water.Therefore,to further improve the robustness of water information extraction algorithms and increase the segmentation precision by extracting more deep information from remote sensing images,this study proposed a water classification method using the semantic segmentation model based on deep learning.First,deep neural networks were used to mine the information from high-resolution remote sensing images.Then,attention modules were used to integrate the deep information with the shallow features such as shape,structure,texture,and hue.Based on the integrated information,a new deep semantic segmentation model with higher precision and prediction efficiency than existent models was built.Finally,the ablation experiment was conducted to compare with conventional recognition methods and common semantic segmentation models.The experiment demonstrates that the proposed algorithm model yields higher overall precision and efficiency than previous methods and that the algorithm parameters are easy to set and less human intervention is required in the model.This study proved the accuracy and efficiency of deep learning and attention mechanism on water information extraction from high-resolution remote sensing images.Moreover,this study provided a possible solution for the segmentation of high-resolution remote sensing images using the deep learning method and explored the future prospect of the solution.

关 键 词：语义分割 多尺度 遥感影像 全卷积网络 注意力机制 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]